Techniques for denoising an electromagnetic signal are disclosed. The techniques utilize an antenna, a weight adaptation component, a reservoir computer including a computer interpretable neural network, a delay embedding component, and an output layer computer. The techniques include passively acquiring an electromagnetic signal by the antenna, producing a plurality of reservoir state values by the reservoir computer based on the electromagnetic signal, collecting the plurality of reservoir state values by the delay embedding component into a historical record, determining a plurality of reservoir state value weights by the weight adaptation component based at least in part of the historical record, scaling, by the plurality of reservoir state value weights, to produce a plurality of output values, the plurality of reservoir state values by the output layer computer, and outputting the plurality of output values, where the scaling occurs over a plurality of clock cycles of a clock for the system.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
2. The system of claim 1, where the reservoir computer, the delay embedding component, the weight adaptation component, and the output layer computer are implemented in at least one Field Programmable gate Array (FPGA).
This invention relates to a reservoir computing system implemented in a Field Programmable Gate Array (FPGA) for processing temporal data. Reservoir computing is a computational paradigm that leverages a fixed, recurrent neural network (the reservoir) to process sequential or time-series data, with only the output layer being trained. The system includes a reservoir computer that processes input data through a recurrent network, a delay embedding component that transforms the input data into a higher-dimensional space to capture temporal dependencies, a weight adaptation component that adjusts the output layer weights based on training data, and an output layer computer that generates predictions or classifications. The FPGA implementation ensures high-speed, parallel processing of the reservoir computing tasks, enabling real-time or near-real-time data analysis. The system is particularly useful for applications requiring fast, adaptive processing of time-dependent signals, such as speech recognition, financial forecasting, or sensor data analysis. By integrating all components into an FPGA, the system achieves low-latency, energy-efficient computation while maintaining the flexibility of reconfigurable hardware. The delay embedding component enhances the system's ability to handle complex temporal patterns, while the weight adaptation component allows for dynamic learning without modifying the reservoir structure. This FPGA-based approach provides a balance between performance and adaptability, making it suitable for embedded or edge computing environments.
3. The system of claim 1, where the reservoir computer, the delay embedding component, the weight adaptation component, and the output layer computer are implemented in at least one Complementary Metal Oxide Semiconductor (CMOS).
A system for reservoir computing is implemented entirely in Complementary Metal Oxide Semiconductor (CMOS) technology. Reservoir computing is a computational paradigm that leverages the dynamic properties of a recurrent neural network, known as the reservoir, to process temporal data. The system addresses the challenge of efficiently implementing reservoir computing in hardware, particularly for real-time applications where low power consumption and high-speed processing are critical. The system includes a reservoir computer, a delay embedding component, a weight adaptation component, and an output layer computer, all fabricated using CMOS. The reservoir computer processes input data by transforming it into a high-dimensional state space, where temporal patterns are more easily separable. The delay embedding component preprocesses the input data to create a time-delayed representation, enhancing the reservoir's ability to capture temporal dependencies. The weight adaptation component adjusts the weights of the connections between the reservoir and the output layer to optimize performance. The output layer computer generates the final output based on the processed data from the reservoir. By integrating all components in CMOS, the system achieves energy efficiency and scalability, making it suitable for edge computing and embedded applications. The use of CMOS also ensures compatibility with existing semiconductor manufacturing processes, reducing development costs and time-to-market. This implementation enables real-time processing of time-series data with low latency, addressing the need for efficient hardware solutions in reservoir computing.
4. The system of claim 1, where the reservoir computer includes electronic memory for storing a reservoir connectivity matrix that models feedback connections between nodes of the neural network.
A reservoir computing system is designed to process complex data streams by leveraging a neural network with dynamic, interconnected nodes. The system addresses challenges in handling non-linear, time-dependent data by using a reservoir of nodes that interact in a structured manner. The reservoir computer includes electronic memory that stores a reservoir connectivity matrix, which defines the feedback connections between the nodes. This matrix models how signals propagate through the network, enabling the system to capture temporal patterns and adapt to varying input conditions. The connectivity matrix can be pre-configured or dynamically adjusted to optimize performance for specific tasks, such as time-series prediction, signal filtering, or pattern recognition. By storing this matrix in electronic memory, the system ensures efficient access and modification of the node interactions, enhancing computational flexibility and scalability. The reservoir computer operates by feeding input data into the network, where the nodes process the signals based on their connections, and the resulting outputs are used for further analysis or control applications. This approach improves processing efficiency and accuracy compared to traditional neural networks, particularly for tasks requiring real-time adaptation.
5. The system of claim 1, where the reservoir computer includes electronic memory for storing feed forward connections from inputs to a plurality of reservoir states.
A system for reservoir computing is disclosed, addressing the challenge of efficiently processing complex, time-dependent data streams. Reservoir computing is a computational paradigm that leverages the dynamic properties of a recurrent neural network, known as a reservoir, to transform input signals into a higher-dimensional state space where subsequent processing, such as linear regression, can be applied. The system includes a reservoir computer with electronic memory for storing feedforward connections from inputs to a plurality of reservoir states. These feedforward connections define how input signals are projected into the reservoir, enabling the system to capture temporal dependencies in the data. The reservoir states, which are dynamically updated based on the input and the reservoir's recurrent connections, form a high-dimensional representation of the input stream. This representation is then used by a readout mechanism, such as a linear classifier or regressor, to produce the desired output. The system is particularly useful for tasks like time-series prediction, signal processing, and control systems where traditional neural networks may struggle due to the complexity of the temporal dynamics. The electronic memory for storing feedforward connections ensures that the system can be configured and adapted for different applications, providing flexibility in modeling various types of input data.
6. The system of claim 1, where the output layer computer is configured to store intermediate values while it scales the plurality of reservoir state values by the plurality of reservoir state value weights over a plurality of clock cycles of a clock for the system.
This invention relates to a neural network system, specifically a reservoir computing system, which processes data by scaling reservoir state values with corresponding weights. The system addresses the challenge of efficiently managing intermediate computations in reservoir computing architectures, where large-scale parallel processing of state values is required. The system includes an output layer computer that stores intermediate values during the scaling process. These intermediate values are generated as the output layer computer scales a plurality of reservoir state values by their respective weights over multiple clock cycles of a system clock. This approach optimizes computational efficiency by retaining intermediate results, reducing redundant calculations, and improving overall processing speed. The system ensures accurate and timely scaling of reservoir state values, which is critical for real-time applications such as time-series prediction, signal processing, and adaptive control systems. By storing intermediate values, the system minimizes memory access delays and enhances throughput, making it suitable for high-performance computing environments. The invention improves upon traditional reservoir computing methods by integrating intermediate storage mechanisms, thereby enhancing computational accuracy and efficiency.
7. The system of claim 1, where the output layer computer is configured to scale the plurality of reservoir state values by the plurality of reservoir state value weights using a cascade of elementary functions.
This invention relates to a neural network system, specifically a reservoir computing system, which processes input data through a reservoir of interconnected nodes to generate reservoir state values. The system addresses the challenge of efficiently transforming these state values into meaningful outputs by using a scalable and flexible output layer. The output layer computer applies a cascade of elementary functions to scale the reservoir state values by corresponding weights, enabling precise and adaptable output generation. The cascade of elementary functions allows for complex transformations while maintaining computational efficiency. This approach enhances the system's ability to model nonlinear relationships in the data, improving accuracy and adaptability in tasks such as time-series prediction, pattern recognition, and control systems. The weighted scaling ensures that the contributions of each reservoir state value are appropriately adjusted, optimizing the system's performance. The invention provides a robust method for refining reservoir computing outputs, making it suitable for applications requiring high precision and dynamic response.
8. The system of claim 1, where each output is produced within Nτmul+log4 N+log4 K clock cycles, where Nτmul is a number of clock cycles used for elementwise pipeline multiplication, N is a size of the reservoir computer, and K is a delay embedding factor.
The invention relates to a reservoir computing system designed to process data with improved computational efficiency. Reservoir computing is a machine learning approach that uses a recurrent neural network with a fixed, randomly connected hidden layer (the reservoir) to process temporal data. The problem addressed is the computational latency in generating outputs from such systems, particularly in applications requiring real-time or near-real-time processing. The system includes a reservoir computer with a configurable size N and a delay embedding factor K, which determines the number of past inputs used to influence the current state. The system performs elementwise multiplication operations in a pipelined manner, reducing the time required for these computations. The key innovation is the optimization of the clock cycle count for producing each output, which is bounded by the expression Nτmul + log4 N + log4 K. Here, Nτmul represents the clock cycles needed for the elementwise multiplication operations, while the logarithmic terms account for the time required to propagate signals through the reservoir and delay embedding layers. By structuring the system to minimize these clock cycles, the invention enables faster output generation without sacrificing accuracy, making it suitable for time-sensitive applications such as signal processing, control systems, and real-time analytics. The design ensures that the computational overhead scales efficiently with the reservoir size and embedding factor, maintaining performance as the system parameters increase.
9. The system of claim 1, where the weight adaptation component is configured to determine the plurality of reservoir state value weights over a plurality of clock cycles of a clock for the system.
The system relates to adaptive control of reservoir computing systems, which are used for processing temporal data. Reservoir computing systems often struggle with dynamically adjusting to changing input patterns, leading to suboptimal performance in real-time applications. The invention addresses this by introducing a weight adaptation component that dynamically adjusts the weights assigned to reservoir state values over multiple clock cycles. This allows the system to better adapt to varying input conditions and improve computational accuracy. The reservoir computing system includes a reservoir layer that processes input data and generates state values, which are then weighted and combined to produce an output. The weight adaptation component continuously evaluates and updates these weights based on system performance metrics, such as error rates or convergence speed, over successive clock cycles. By dynamically adjusting the weights, the system can maintain high accuracy even as input patterns evolve. This adaptive mechanism enhances the system's ability to handle time-varying data, making it suitable for applications like signal processing, time-series forecasting, and control systems where real-time adaptability is critical. The invention improves upon static weight configurations by incorporating a feedback loop that refines weights in response to ongoing system behavior.
10. The system of claim 1, where the system includes a single antenna.
A wireless communication system with a single antenna is designed to optimize signal transmission and reception in environments where space or hardware constraints limit the use of multiple antennas. The system addresses challenges such as signal interference, limited bandwidth, and reduced efficiency in conventional single-antenna setups. By integrating advanced signal processing techniques, the system enhances performance by dynamically adjusting transmission parameters, such as power levels and modulation schemes, to improve reliability and data throughput. The system may also incorporate beamforming or directional control to focus signal energy in specific directions, compensating for the lack of multiple antennas. Additionally, the system may include adaptive algorithms to mitigate multipath fading and interference, ensuring stable communication even in challenging conditions. The single-antenna design simplifies deployment while maintaining competitive performance, making it suitable for applications in IoT devices, portable communication systems, and other space-constrained environments. The system may further include error correction and signal amplification features to enhance robustness against noise and signal degradation. Overall, the invention provides a cost-effective and efficient solution for wireless communication in scenarios where multiple antennas are impractical.
12. The method of claim 11, where the reservoir computer, the delay embedding component, the weight adaptation component, and the output layer computer are implemented in at least one Field Programmable gate Array (FPGA).
This invention relates to a reservoir computing system implemented on a Field Programmable Gate Array (FPGA) for processing time-series data. The system addresses the challenge of efficiently implementing reservoir computing architectures in hardware, which traditionally require significant computational resources and power consumption. The invention combines a reservoir computer, a delay embedding component, a weight adaptation component, and an output layer computer, all integrated into an FPGA to enable real-time processing of dynamic data streams. The reservoir computer generates high-dimensional representations of input data, while the delay embedding component transforms the input data into a form suitable for reservoir processing. The weight adaptation component adjusts the system's parameters to optimize performance, and the output layer computer produces the final output based on the processed data. By implementing these components on an FPGA, the system achieves low-latency, energy-efficient computation, making it suitable for applications in signal processing, predictive modeling, and control systems where real-time performance is critical. The FPGA-based implementation also allows for flexibility in reconfiguring the system for different tasks without requiring hardware redesign.
13. The method of claim 11, where the reservoir computer, the delay embedding component, the weight adaptation component, and the output layer computer are implemented in at least one Complementary Metal Oxide Semiconductor (CMOS).
This invention relates to a reservoir computing system implemented in Complementary Metal Oxide Semiconductor (CMOS) technology. Reservoir computing is a neural network approach that processes temporal data by leveraging a fixed, recurrent neural network (the reservoir) to project input signals into a higher-dimensional space, followed by a trainable output layer for classification or regression tasks. The system addresses challenges in hardware implementation, such as power efficiency, scalability, and real-time processing, by integrating key components into a CMOS architecture. The system includes a reservoir computer that processes input signals through a recurrent network with fixed weights, generating a high-dimensional state representation. A delay embedding component transforms the input data into a format suitable for the reservoir, ensuring temporal dependencies are preserved. A weight adaptation component adjusts the output layer weights based on training data, enabling the system to learn and generalize from input patterns. The output layer computer processes the reservoir's state to produce final predictions or classifications. By implementing these components in CMOS, the system achieves low-power operation, compact form factors, and compatibility with existing semiconductor manufacturing processes. This approach is particularly useful for edge computing applications where real-time processing of time-series data, such as sensor signals or financial time series, is required. The CMOS-based design ensures energy efficiency and scalability, making it suitable for deployment in resource-constrained environments.
14. The method of claim 11, where the producing a plurality of reservoir state values is performed using an electronically stored reservoir connectivity matrix that models feedback connections between nodes of the neural network.
This invention relates to neural network modeling, specifically improving the accuracy of reservoir computing systems by incorporating feedback connections between nodes. Reservoir computing is a type of recurrent neural network where a fixed, randomly connected "reservoir" of nodes processes input data, and a readout layer extracts useful information. A key challenge in reservoir computing is ensuring the reservoir's dynamics effectively capture the temporal patterns in the input data, which often requires careful tuning of the reservoir's structure. The invention addresses this by using an electronically stored reservoir connectivity matrix to model feedback connections between nodes in the neural network. This matrix defines how signals propagate within the reservoir, allowing for controlled feedback loops that enhance the network's ability to retain and process temporal information. By adjusting the connectivity matrix, the system can optimize the reservoir's state dynamics to better represent the underlying data patterns. The method involves producing a plurality of reservoir state values based on these feedback connections, which are then used to generate output predictions or classifications. This approach improves the network's performance in tasks requiring temporal sequence processing, such as time-series forecasting or speech recognition, by enabling more sophisticated interactions between nodes. The connectivity matrix can be pre-trained or dynamically adjusted during operation to adapt to different input data characteristics.
15. The method of claim 11, where the producing a plurality of reservoir state values is performed using electronically stored feed forward connections from inputs to a plurality of reservoir states.
This invention relates to neural network systems, specifically reservoir computing, which processes temporal data by mapping inputs to reservoir states. The problem addressed is the need for efficient and accurate state generation in reservoir computing systems to improve temporal data processing tasks like prediction, classification, or control. The method involves producing a plurality of reservoir state values by leveraging electronically stored feed-forward connections from input data to multiple reservoir states. These connections define how input signals propagate through the reservoir, generating dynamic state representations. The reservoir states are then used to compute output values, which can be further processed for tasks like prediction or classification. The feed-forward connections ensure that the reservoir states evolve in a structured manner, enhancing the system's ability to capture temporal dependencies in the input data. The method may also include training the feed-forward connections to optimize performance for specific tasks, such as minimizing prediction errors or improving classification accuracy. This approach improves the efficiency and accuracy of reservoir computing systems by systematically mapping inputs to reservoir states using predefined connections.
16. The method of claim 11, where scaling includes storing intermediate values while the output layer computer scales the plurality of reservoir state values by the plurality of reservoir state value weights over a plurality of clock cycles of a clock for the system.
This invention relates to a method for scaling reservoir state values in a neural network system, particularly in a spiking neural network (SNN) or similar architecture. The problem addressed is the efficient scaling of reservoir state values, which are intermediate outputs from a reservoir computing layer, to produce final outputs in a computationally optimized manner. The method involves scaling a plurality of reservoir state values by corresponding reservoir state value weights to generate a plurality of scaled reservoir state values. This scaling process is performed by an output layer computer, which processes the reservoir state values over multiple clock cycles of a system clock. During this scaling, intermediate values are stored to facilitate the computation. The reservoir state values are typically generated by a reservoir computing layer, which processes input data in a recurrent or dynamic manner, capturing temporal patterns. The output layer computer then applies the scaling weights to these values, which may be stored in memory or registers, to produce the final scaled outputs. The method ensures that the scaling operation is performed efficiently, leveraging the system's clock cycles to handle the computational load without requiring excessive hardware resources. This approach is particularly useful in real-time or resource-constrained environments where processing speed and memory efficiency are critical. The stored intermediate values allow for pipelined or staged computation, improving throughput and reducing latency.
17. The method of claim 11, where the scaling includes using a cascade of elementary functions.
A method for scaling data in computational systems addresses the challenge of efficiently processing large datasets while maintaining accuracy and computational efficiency. The method involves applying a cascade of elementary functions to scale the data. Elementary functions are basic mathematical operations, such as linear transformations, polynomial functions, or exponential functions, that are computationally simple yet effective for scaling purposes. By arranging these functions in a cascade, where the output of one function serves as the input to the next, the method achieves a more refined and precise scaling process. This approach allows for fine-tuning the scaling parameters to better adapt to the characteristics of the input data, improving the overall performance of data processing tasks. The cascade structure also enables modularity, allowing individual functions to be adjusted or replaced without disrupting the entire scaling process. This method is particularly useful in applications requiring high precision, such as machine learning, signal processing, and data normalization, where accurate scaling is critical for subsequent analysis or modeling. The use of elementary functions ensures that the scaling process remains computationally efficient while achieving the desired accuracy.
18. The method of claim 11, where each output is produced within Nτmul+log4 N+log4 K clock cycles, where Nτmul is a number of clock cycles used for elementwise pipeline multiplication, N is a size of the reservoir computer, and K is a delay embedding factor.
This invention relates to reservoir computing, a computational paradigm that processes temporal data by leveraging a recurrent neural network with a fixed, randomly connected structure. The problem addressed is the computational efficiency of reservoir computing systems, particularly in reducing the time required to produce outputs while maintaining accuracy. The method involves a reservoir computer with a fixed-size reservoir and a delay embedding factor. The reservoir processes input data through a series of operations, including elementwise multiplication, to generate outputs. A key aspect is the optimization of the computational pipeline to ensure that each output is produced within a bounded number of clock cycles, specifically Nτmul + log4 N + log4 K, where Nτmul represents the clock cycles needed for elementwise pipeline multiplication, N is the reservoir size, and K is the delay embedding factor. This ensures predictable and efficient processing times, which is critical for real-time applications. The method also includes steps for initializing the reservoir, applying input data, and generating outputs through recurrent connections. The fixed structure of the reservoir allows for parallel processing, further improving efficiency. The delay embedding factor K determines the temporal depth of the input data, enabling the system to capture long-term dependencies. The logarithmic terms in the clock cycle count reflect the efficiency of the computational steps, such as matrix operations, which scale logarithmically with reservoir size and delay embedding factor. This approach balances computational speed with the ability to handle complex temporal patterns.
19. The method of claim 11, where the determining a plurality of reservoir state value weights occurs over a plurality of clock cycles of a clock for the system.
This invention relates to a method for determining reservoir state value weights in a computational system, particularly in the context of reservoir computing or similar dynamic systems. The method addresses the challenge of efficiently calculating and updating these weights, which are critical for the system's performance in tasks such as time-series prediction, pattern recognition, or control applications. The weights influence how the system processes and retains information from its reservoir, a key component in reservoir computing that captures temporal dynamics. The method involves distributing the computation of reservoir state value weights across multiple clock cycles of the system's clock. This approach allows for more efficient processing, especially in hardware implementations where parallel or pipelined operations are beneficial. By spreading the computation over time, the method reduces the computational load per cycle, enabling real-time or near-real-time operation. The weights are derived from the reservoir's state values, which are typically generated by passing input data through a recurrent or dynamic network. The method may also involve adjusting these weights based on feedback or error signals to improve system accuracy. The technique is particularly useful in systems where hardware constraints limit the number of operations that can be performed in a single clock cycle. By leveraging multiple clock cycles, the method ensures that the system can handle complex computations without requiring excessive hardware resources. This approach is applicable in various domains, including neural networks, signal processing, and control systems, where dynamic adaptation and efficient computation are essential.
20. The method of claim 11, where the system includes a single antenna.
A system and method for wireless communication using a single antenna to transmit and receive signals in a shared frequency band. The system addresses challenges in wireless communication where multiple devices operate in the same frequency spectrum, leading to interference and reduced performance. The method involves dynamically adjusting transmission parameters, such as power, timing, or modulation, to minimize interference with other devices while maintaining reliable communication. The system may also include a controller that monitors signal quality and adjusts parameters in real-time to optimize performance. The single antenna design simplifies hardware complexity and reduces cost compared to multi-antenna systems, making it suitable for compact or low-power devices. The method further includes techniques for detecting and avoiding collisions with other transmissions, ensuring efficient use of the shared spectrum. By dynamically adapting to environmental conditions, the system improves communication reliability and throughput in congested wireless environments. The approach is particularly useful in applications where spectrum efficiency and hardware simplicity are critical, such as IoT devices, sensor networks, or mobile communication systems.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
July 26, 2018
December 20, 2022
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.